The software path
A version printed at one layer cannot by itself establish the version or support state of every other layer.
[5][33][36][37]The physical integration path
An application platform sits above the hardware map
NVIDIA Omniverse supplies OpenUSD-based building blocks for 3D and physical-AI applications. It answers how a team can exchange scene data, build simulation tools, and deploy an experience; it is not another GPU, HGX baseboard, or DGX system.
[90][91]NVLink, Ethernet, and InfiniBand answer different connection questions
NVLink is a GPU interconnect inside supported systems or board configurations. Spectrum-X is NVIDIA’s Ethernet platform for AI cluster networking. Quantum InfiniBand is its InfiniBand platform. NVIDIA’s networking overview separates NVLink scale-up from Ethernet or InfiniBand scale-out; a GPU interconnect does not replace the cluster network.
If you are selecting a cluster fabric, start with the network and system design, then check the exact switch, adapter, server, and software support. If you are checking whether two graphics cards can be bridged, start with the exact card specifications instead: NVIDIA lists NVLink (SLI-Ready) as No for GeForce RTX 4090 and RTX 5090 reference products. Do not infer support for another RTX or professional card from its family name.
[96][97][98][99][100]Use the map to diagnose the claim
If a press release says a platform is “available,” ask: which GPU or system, through which provider or channel, in which region, with which software version? If a workload fails, ask: is the issue hardware feature support, driver capability, toolkit installation, a library matrix, or application packaging?
The map turns a vague product question into a sequence of checkable claims.
[23][33][31]